Integration of statistical collocation segmentations in a phrase-based statistical machine translation system

Integration of statistical collocation segmentations in a phrase-based statistical machine translation system Marta R. Costa-juss`a∗, Vidas Daudaravicius† and Rafael E. Banchs∗ ∗Barcelona Media Research Center Av Diagonal, 177, 9th floor, 08018 Barcelona, Spain {marta.ruiz,rafael.banchs}@barcelonamedia.org † Faculty of Informatics, Vytautas Magnus University Vileikos 8, Kaunas, Lithuania vidas@donelaitis.vdu.lt Abstract This study evaluates the impact of inte- grating two different collocation segmen- tations methods in a standard phrase-based statistical machine translation approach. The collocation segmentation techniques are implemented simultaneously in the source and target side. Each resulting col- location segmentation is used to extract translation units. Experiments are reported in the English-to-Spanish Bible task and promising results (an improvement over 0.7 BLEU absolute) are achieved in trans- lation quality. 1 Introduction Machine Translation (MT) investigates the use of computer software to translate text or speech from one language to another. Statistical machine trans- lation (SMT) has become one of the most popu- lar MT approaches given the combination of sev- eral factors. Among them, it is relatively straight- forward to build an SMT system given the freely available software and, additionally, the system construction does not require of any language ex- perts. Nowadays, one of the most popular SMT ap- proaches is the phrase-based system (Koehn et al., 2003) which implements a maximum entropy ap- proach based on a combination of feature func- tions. The Moses system (Koehn et al., 2007) is an implementation of this phrase-based ma- chine translation approach. An input sentence is first split into sequences of words (so-called phrases), which are then mapped one-to-one to tar- get phrases using a large phrase translation table. c⃝ 2010 European Association for Machine Translation. Introducing chunking in the standard phrase- based SMT system is a relatively frequent study (Zhou et al., 2004; Wang et al., 2002; Ma et al., 2007). Chunking may be used either to improve reordering or to enhance the translation table. For example, authors in (Zhang et al., 2007) present a shallow chunking based on syn- tactic information and they use the chunks to re- order phrases. Other studies report the impact on the quality of word alignment and in translation after using various types of multi-word expres- sions which can be regarded as a type of chunks, see (Lambert and Banchs, 2006) or sub-sentential sequences (Macken et al., 2008; Groves and Way, 2005). Chunking is usually performed on a syn- tactic or semantic basis which forces to have a tool for parsing or similar. We propose to introduce the collocation segmentation developed by (Dau- daravicius, 2009) which is language independent. This collocation segmentation was applied in key- word assigment task and a high classification im- provement was achieved (Daudaravicius, 2010). We use this collocation segmentation technique to enrich the phrase translation table. The phrase translation table is composed of phrase units which generally are extracted from a word aligned paral- lel corpus. Given this word alignment, an extrac- tion of contiguous phrases is carried out (Zens et al., 2002), specifically all extracted phrases fulfill the following restrictions: all …

Content truncated to 500 words